While generative AI in higher education often brings to mind assistance (authorized or not) in achieving a final product or saving time at the cost of learning, many instructors are exploring the potential for using generative AI to create resources that support students in specific areas of need. A common (and valid) question comes up in discussions of AI’s place in higher education: “What problem was this solving?” or “Why did we need this?”

This blog post highlights the work of Selma Yildirim, a UChicago math instructor whose use case for these tools offers an answer to this question. She noticed an area in which her students genuinely needed a new resource, and used the university’s dedicated AI tool, PhoenixAI, to build it. While services like ChatGPT offer the ability to create custom chatbots, she found that PhoenixAI offered an experience that was accessible to all students, regardless of the resources they had available to pay for premium services (or comfort creating an account with a commercial tool). In fact, without PhoenixAI, Yildirim explains, she would likely not have pursued the project:

“My main motivation for using PhoenixAI was that everyone can access it,” Yildirim explains. “It is a little bit more structured and limited, and it is only for the UChicago community. I wouldn’t tell students to use ChatGPT because they may not have an account or some of them may have a premium account while some of them don’t and are using the free version, so I just wanted to use PhoenixAI. So if it wasn’t PhoenixAI, I would probably not be using ChatGPT.”

What need did this fill?

While teaching calculus, Yildirim, an Associate Instructional Professor in the Department of Mathematics, noticed a need: students were coming into her calculus classes with some gaps in their foundational knowledge from precalculus and even trigonometry. While she herself offers support to her students as well as sessions with junior tutors, that limited time is best used to support students in the calculus concepts of the current class.

As a result, Yildirim saw the need not for an all-purpose math tool, but for a specifically designed supplemental tool to assist students needing to develop or review various precalculus foundations in order to thrive in a calculus class.

According to Yildirim, “I always liken this to a telescope: when they want it, it opens up more and more. Not all of our students need the same thing. So if they need it, the help is there. It is more dynamic, not static.”

Rather than programming the tool, “Math 131 – Precalculus AI Assistant” (now on v.2) to cover everything in the precalculus course, Yildirim designed the core of her chatbot to cover specific material from more basic coursework that students need for her calculus course, complete with a curated problem set, topic lists, and a routine to respond to student answers by going through the solution steps and confirm the answer. True to Yildirim’s telescope analogy, she’s designed another of her tools, the Precalculus AI chatbot, to give users options to extend their session with either additional practice or review of a new topic. For example, after reviewing basics of lines it ends the interaction with the question like:

“Would you like to learn more about the topic of lines, or do you prefer to review another topic or practice some questions?”

In some ways, the creation of this tool is not unlike the creation of non-AI resources we’ve highlighted on the ATS blog before, like Aidan Kaplan’s supplemental grammar videos to help get students up to speed in concepts they may need for his Arabic classes, but which are technically foundational concepts from previous classes. Like Kaplan’s bite-sized videos to help with specific tenses that students can access as needed, Yildirim’s chatbots help novice learners access information to catch up with the benefit of expert design. In this case, though, Yildirim takes advantage of the conversational fluency of AI tools while also pointing students to helpful information and practice exercises.

Additionally, Yildirim explains that one of the reasons she’s built subject-specific chatbots is to shape the AI use that already happens among students in a positive direction: “That was one of the main reasons for me building this, because students are using AI. Meeting students where they are, providing more aligned content…because they are using it. So as an instructor, this is me helping them to find—to shape that a little bit…without overwhelming them.”

How do you design this?

When Yildirim started to develop custom GPTs like her precalculus helper, she had a few principles in mind as she designed the student experience:

• Simplicity
• Flexibility
• Ease of use

That meant not just providing students with information and her own practice problems, but also the navigational help to find what they need. Yildirim has designed her tool to guide the user’s path to specific categories of help–a useful function when you don’t know “what you don’t know!”

From Yildirim’s perspective, it’s important to provide a more structured experience than a standard, general-purpose AI tool would. To that end, she designs her chatbots to help students figure out what kind of support they need. As she explains it: “Students don’t know exactly their level because they are still learning. Alignment, I think, is the key word here.”

With that in mind, the user experience for her Precalculus AI Assistant starts with three initial options curated by Yildirim:

• Review of Precalculus Topics
• Practice Questions
• History of Math

By using AI, she takes advantage of the fluency with language and dynamic conversational nature that’s made these tools so engaging to users in other contexts. Based on the user’s answer, the chatbot continues to ask questions about what specific knowledge they’re looking for, giving options, background information, and examples–following an expert-designed path to find the right information at the right time. For example, when they ask for a review of precalculus topics, the chatbot offers them a list of options curated by Yildirim herself:

1. Cartesian Plane and Functions
2. Lines
3. Polynomials and Rational Functions
4. Inequalities
5. Exponential and Logarithmic Functions
6. Trigonometry

To encourage the user to engage with concepts and practice skills they need to succeed in calculus, Yildirim’s chatbot combines a few strengths:

• The conversational fluency of AI: it’s easy to talk to and it phrases explanations in approachable language.

• Content determined by the educator: Having taught students in this subject and knowing potential pitfalls, Yildirim knew what kind resources to make available.

• Ease of navigation: By writing detailed instructions for the AI, Yildirim made sure it offers students a list of topics appropriate to their level and that it always offers them options to dig in further.

She describes this part of her design as a way of both encouraging them to study more and to give them the resources they need without having to look up additional information elsewhere:

“it gives these fundamental explanations and asks ‘would you want more details, to review another topic, or practice?’ The ease of use means they can navigate without leaving [the tool] and they can review or practice more.”

While the next section will introduce readers to the technical resources and how-to of setting up something like this, it’s important to note that the reflection and vision that goes into design was a major part of Yildirim’s work. It requires thinking through instructions for the chatbot to interact with students in a certain way, giving it resources to answer as desired, testing, and revising the directions. (For an example from another discipline of the kind of testing and iteration this requires, check out Cornell University’s podcast episode on building chatbots for students of German.)

Create a Chatbot: Higher-Level Considerations

1. 1. Reflect on user path
      After identifying the knowledge that her students needed, Yildirim took some time to sketch out the path of the user as she was writing instructions for her chatbot. By listing the topics they might need to explore and thinking through different ways into the subject matter, she was able to write better instructions, creating a smoother experience for the students. There are those who’ve referred to AI prompting as “coding in prose,” which may be helpful as you consider the specific and concise instructions you build into your custom chatbot.
   2. Curate your resources
      What do you want to make sure your students have access to? Knowing the resources that matter to you can help you think about what you want to direct the AI to reference during the user experience.
   3. Write your instructions
      Tell the tool what you want it to do. If you’re having a hard time figuring out where to start, keep it simple to start. Think about what you write in three categories:
      Task: Be specific about what you want the AI to do.
      Instructions: What type of output do you want?, are there steps you want it to follow?; do you want it to assume a particular persona? (A good example of the customization here is that Yildirim instructs one of her tools to provide hints rather than examples for practice problems; it’s more learning-friendly, and it helps to avoid providing misinformation!)
      Context: Is there background info the tool needs to do the task well? Is there a set of questions you want it to ask the user in order to do its work?
      (At the end of this post, you’ll find a list of LinkedIn Learning courses Yildirim found useful in picking up the skills she needed to write good instructions.)
   4. Test, iterate, and test again

Yildirim stresses how important it was to test her tool repeatedly before she released it for student use, adjusting the directions for the chatbot as she encountered various functions she wanted to adjust.

Additionally, Yildirim herself emphasizes the need to update these chatbots on an ongoing basis in response to new knowledge, module updates, and student feedback. These designs should evolve based on the lessons learned from use.

Create a Chatbot: The Technical Steps

To access PhoenixAI, go to https://phoenixai.uchicago.edu and log in using your UChicago credentials. Click “Create an Assistant” in the top right corner.

The basic components you need to fill in to set up an Assistant are:

• • Name: What will your chatbot be called?
  • Description: What do you want users to know about this tool before they start using it?
  • Instructions: This is the most important section. By giving the tool instructions to follow in every interaction, you can make sure it follows a sequence of steps, references specific files, offers specific kinds of information, uses a specific voice, and even avoids certain kinds of behavior. Think back to the examples of what Yildirim’s chatbots do to get an idea of what you can prompt:
    • • One of Yildirim’s chatbots (about practicing some topics) gives a specific set of topic and difficulty level options pulled from one of her documents.
    • When a student asks for practice problems, it has instructions to use the ones Yildirim has provided.
    • In another example, her instructions to her “Precalculus AI Assistant” tell the AI to explain how any given concept they’re working on will apply to their work in differential calculus, addressing the relevance of the outputs. It even provides notes on practical applications of the concepts.
    • And crucially, although she’s uploaded reliable documentation to mitigate inaccurate outputs from AI, she’s also given the tool directions to consistently remind students that these tools are still always capable of making mistakes. This is a vital piece of information to keep in students’ minds, as AI features will show up in more tools they use, sometimes without being easily perceived as AI outputs that need a critical eye.
  • Conversation starters: Initial messages users can click to start a conversation. You can use these, as Yildirim does, to prompt students to start their conversations on a particular set of suggested paths.
  • Knowledge: This is where you can upload specific information or domain-specific data to specialize your GPT for a particular task or use case. By curating and uploading factual, high-quality sources, you mitigate the risk of inaccurate outputs, while allowing the user to interact with that body of knowledge through a fluent interface. You can also build in specific directions to the chatbot to reference a given document.
    • Supported formats are: c, cs, cpp, doc, docx, html, java, json, md, pdf, php, pptx, py, rb, tex, txt, css, js, sh, ts
    • (You may upload up to 10 files at once. Maximum number of files per Assistant is 100 files. Upload is limited to 20MB per file.)
  • Shared or private?
    • When ready, you can share a link to use your chatbot with students or other users.

• Enable Code Interpreter?: Code Interpreter enables the Assistant to read and analyze file uploads, write and execute computer code, generate files, and create visualizations from data. This can be handy if you want students to upload files for the tool to interact with or if you’d like users to be able to download certain kinds of outputs.

Looking to Build Knowledge?

LinkedIn Learning

Based on her extensive experimentation building chatbots and her use of LinkedIn Learning to develop her design skills, Professor Yildirim has curated a Learning Path in LinkedIn Learning to share with interested colleagues, “Building Custom GPTs for Teaching & Learning.” This learning path, a playlist of sorts for self-directed AI skills development, contains courses that Yildirim found helpful as she acquired the skills and the language she needed to build her chatbots. The Learning Path includes courses such as:

• Introduction to Prompt Engineering for Generative AI
• OpenAI ChatGPT: Creating Custom GPTs
• ChatGPT: Crafting Exceptional GPTs for Enhanced Productivity and Innovation

Remember, all UChicago students, staff, and instructors have access to LinkedIn Learning using their UChicago CNet ID.

Teaching in the Generative AI Landscape Canvas Course

Academic Technology Solutions (ATS), the Chicago Center for Teaching and Learning (CCTL), and the Library have created a single resource where you can find guidance and resources on generative AI. Whatever your teaching context or disposition toward AI, you can find useful resources in the new Canvas site Teaching in the Generative AI Landscape. Browse the self-guided course or check out a preview on our blog.

Custom Chatbots in PhoenixAI

To learn more about building a custom AI using PhoenixAI Assistant, visit the Assistant Knowledge Base article and FAQ. PhoenixAI is still relatively new to UChicago, so please reach out to ATS with any questions about how this tool might support your work.

Further Information

To assess the purposes these tools might serve in your course and the best structure to support thoughtful use, check out our blog post presenting a framework for these decisions. To start by acquiring comfort with prompt writing for GAI, check out our post with guidance and sample prompts. Subscribe to ATS’ blog and newsletter for updates on new resources to support you in using these tools.

For individual assistance, you can visit our office hours, book a consultation with an instructional designer, or email academictech@uchicago.edu. For a list of our upcoming ATS workshops, please visit our workshop schedule for events that fit your schedule.

Header Image by sandid from Pixabay